Background/Objectives: The purpose of this study was to analyze relevant areas in acute-stage fluorescein angiography (FA) images, predicting the long-term visual prognosis of branch retinal vein occlusion (BRVO) based on gradient-weighted class activation mapping (Grad-CAM). Methods: This retrospective observational study included 136 eyes with BRVO that were followed up for more than a year post-FA. Cropped grayscale images centered on the fovea (200 × 200 pixels) were manually pre-processed from early-phase FA at the acute phase. Pairs of the cropped FA images and the best-corrected visual acuity (BCVA) in remission at least one year post-FA were used to train a 38-layer ResNet with five-fold cross-validation. Correlations between the ResNet-predicted and true (actually measured) logMAR BCVAs in remission, and between the foveal avascular zone (FAZ) area measured by ImageJ (version 1.52r) from FA images and true logMAR BCVA in remission were evaluated. The heat maps generated by Grad-CAM were evaluated to determine which areas were consumed as computational resources for BCVA prediction. Results: The correlation coefficient between the predicted and true logMAR BCVAs in remission was 0.47, and that between the acute-stage FAZ area and true logMAR BCVA in remission was 0.42 (p < 0.0001 for both). The Grad-CAM-generated heat maps showed that retinal vessels adjacent to the FAZ and the FAZ per se had high selectivity (95.7% and 62.2%, respectively). Conclusions: The Grad-CAM-based analysis demonstrated FAZ-neighboring vessels as the most relevant predictor for the long-term visual prognosis of BRVO.

Background/Objectives: The aim of this study was to determine artificial intelligence-based macular fluid (MF) parameters in diabetic macular edema (DME) with optical coherence tomography (OCT) and examine stage-by-stage differences in MF parameters and their relationship with best-corrected visual acuity (BCVA). Methods: This study enrolled 104 eyes with treatment-naïve DME. Intraretinal fluid (IRF) and subretinal fluid (SRF) were detected in horizontal OCT images based on the “Hokkaido University MF segmentation model” when DME was first observed together with BCVA testing. The MF area, the mean brightness, and the variance of brightness were compared between mild or moderate non-proliferative diabetic retinopathy (mNPDR, n = 33), severe NPDR (sNPDR, n = 52), and PDR eyes (n = 19). Correlations between logMAR BCVA and MF parameters were also examined. Results: All the MF parameters tended to increase with DR stages. Especially, the mean brightness of IRF was significantly greater in PDR than in mNPDR. The variance of brightness of IRF increased in sNPDR compared to mNPDR, whereas that of SRF increased in PDR compared to sNPDR. LogMAR BCVA showed positive correlations with MF areas and the variance of brightness of SRF. Conclusions: The qualitative and quantitative MF parameters may be useful for better understanding DME pathogenesis according to DR progression.

To compare the prediction accuracy of brachial-ankle pulse wave velocity (baPWV) from color fundus photographs (CFPs) using different deep learning models. This retrospective study analyzed the data of 696 participants whose baPWVs and CFPs were obtained during medical checkups. Arteriolar and venular probability maps, which were automatically calculated from the CFPs based on our modified deep U-net, Hokkaido University retinal vessel segmentation (HURVS) model, were applied as channel attention to retinal vessel location information to predict baPWV. The baPWV prediction parameters consisted of predicted baPWVs from a single-input model using CFPs only and from a three-input model using CFPs, and arteriolar and venular probability maps. The single- and three-input models adopted a common depth-wise net and were separately pretrained and trained with fivefold cross-validation. These baPWV prediction parameters were corrected using multiple regression equations with age, sex, and systolic blood pressure and were defined as single- and three-input regression-predicted baPWVs. The main outcome measures were the correlation coefficients between true baPWV and the baPWV prediction parameters. The correlation coefficient with true baPWVs was higher for the three-input predicted baPWVs (R = 0.538) than for the single-input predicted baPWVs (R = 0.527). After regression, the three-input, regression-predicted baPWVs (R = 0.704) had the highest prediction accuracy, followed by the single-input, regression-predicted baPWVs (R = 0.692). The three-input model predicted true baPWVs with high accuracy. This improved prediction accuracy by channel attention to the arteriolar and venular probability maps based on the HURVS model confirmed that arterioles and venules are relevant regions for baPWV prediction.